Abstract: Provincial or municipal power supply companies need to forecast the distribution network material demand in the new year at the end of the year or at the beginning of the year. The previous estimation method based on experience has poor accuracy and low efficiency. After obtaining the material claim data in the ERP system, the materials are divided into dozens of standard packages according to the use and type, then, dividing the time series of the amount of material claim of each package by the investment amount of the corresponding year, the time series of material claim of the unit investment amount of the monthly granularity can be obtained, classical algorithms such as LSTM long-short memory neural network, Croston, primary exponential smoothing and secondary exponential smoothing are used to predict the sequence. At last, the material demand forecasting result is obtained by using the weighted correction from the series of annual material demand of unit investment. The method proposed in this paper is tested on the data of Zhejiang, Jiangsu, Fujian, Sichuan and Shaoxing, and the results are satisfactory.